qwen3-omni
Mirror of the Qwen3-Omni omni-modal model in Transformers safetensors format.
What it is for
The repo holds a copy of Qwen3-Omni: sharded safetensors weights, the model and generation configs, the chat template, the preprocessor config and the tokenizer.
Qwen3-Omni is an end-to-end multilingual model. It takes text, images, audio and video, and it answers in streamed text and speech.
It uses a mixture-of-experts Thinker-Talker design. FORK_PROVENANCE.md records where the copy came from.
Provenance
- Upstream source: Qwen/Qwen3-Omni-30B-A3B-Instruct.
- Code and cookbooks: QwenLM/Qwen3-Omni.
- Technical report: Qwen3-Omni Technical Report.
Use
hf download chibifire/qwen3-omni --local-dir qwen3-omni
from transformers import Qwen3OmniMoeForConditionalGeneration, Qwen3OmniMoeProcessor
model = Qwen3OmniMoeForConditionalGeneration.from_pretrained("chibifire/qwen3-omni", dtype="auto", device_map="auto")
processor = Qwen3OmniMoeProcessor.from_pretrained("chibifire/qwen3-omni")
See the upstream card for the install steps, vLLM use and voice options.
Licence
Apache 2.0, the licence of the Qwen3-Omni code repo. Terms: LICENSE.